class ORTTransformerTrainer:
"""
"""
model: PreTrainedModel
args: TrainingArguments
train_dataset: Dataset
eval_dataset: Dataset
compute_metrics: Callable[[EvalPrediction], Dict]
def __init__(
self,
model: PreTrainedModel,
model_desc: ModelDescription,
args: TrainingArguments,
train_dataset: Dataset,
eval_dataset: Dataset,
compute_metrics: Callable[[EvalPrediction], Dict],
):
"""
"""
self.model = model
self.model_desc = model_desc
self.args = args
self.data_collator = DefaultDataCollator()
self.train_dataset = train_dataset
self.eval_dataset = eval_dataset
self.compute_metrics = compute_metrics
set_seed(self.args.seed)
# Create output directory if needed
if self.args.local_rank in [-1, 0]:
os.makedirs(self.args.output_dir, exist_ok=True)
def get_train_dataloader(self) -> DataLoader:
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
train_sampler = (SequentialSampler(self.train_dataset)
if self.args.local_rank == -1 else DistributedSampler(
self.train_dataset))
return DataLoader(
self.train_dataset,
batch_size=self.args.train_batch_size,
sampler=train_sampler,
collate_fn=self.data_collator.collate_batch,
)
def get_eval_dataloader(self) -> DataLoader:
return DataLoader(
self.eval_dataset,
batch_size=self.args.eval_batch_size,
shuffle=False,
collate_fn=self.data_collator.collate_batch,
)
def get_test_dataloader(self, test_dataset: Dataset) -> DataLoader:
# We use the same batch_size as for eval.
return DataLoader(
test_dataset,
batch_size=self.args.eval_batch_size,
shuffle=False,
collate_fn=self.data_collator.collate_batch,
)
def train(self):
"""
Main training entry point.
"""
train_dataloader = self.get_train_dataloader()
if self.args.max_steps > 0:
t_total = self.args.max_steps
num_train_epochs = (self.args.max_steps //
(len(train_dataloader) //
self.args.gradient_accumulation_steps) + 1)
else:
t_total = int(
len(train_dataloader) //
self.args.gradient_accumulation_steps *
self.args.num_train_epochs)
num_train_epochs = self.args.num_train_epochs
get_lr_this_step = get_linear_schedule_with_warmup(
self.args.warmup_steps, t_total, self.args.learning_rate)
loss_scaler = LossScaler('loss_scale_input_name',
True,
up_scale_window=2000)
def map_optimizer_attributes(name):
# no_decay_keys = ["bias", "LayerNorm.weight"]
no_decay = "bias" in name or "LayerNorm.weight" in name
if no_decay:
return {"weight_decay": 0.0, "weight_decay_mode": 1}
else:
return {
"weight_decay": self.args.weight_decay,
"weight_decay_mode": 1
}
self.model = ORTTrainer(
self.model,
None,
self.model_desc,
"AdamOptimizer",
map_optimizer_attributes=map_optimizer_attributes,
learning_rate_description=IODescription('Learning_Rate', [
1,
], torch.float32),
device=self.args.device,
gradient_accumulation_steps=self.args.gradient_accumulation_steps,
world_rank=0,
world_size=1, # only support single GPU cases
use_mixed_precision=self.args.fp16,
allreduce_post_accumulation=True,
get_lr_this_step=get_lr_this_step,
loss_scaler=loss_scaler,
enable_grad_norm_clip=False,
_opset_version=12,
_use_deterministic_compute=True)
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataloader.dataset))
logger.info(" Num Epochs = %d", num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
self.args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
self.args.train_batch_size *
self.args.gradient_accumulation_steps *
(torch.distributed.get_world_size()
if self.args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
self.args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
epochs_trained = 0
steps_trained_in_current_epoch = 0
tr_loss = 0.0
logging_loss = 0.0
train_iterator = trange(
epochs_trained,
int(num_train_epochs),
desc="Epoch",
disable=self.args.local_rank not in [-1, 0],
)
for epoch in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=self.args.local_rank not in [-1, 0])
for step, inputs in enumerate(epoch_iterator):
# Skip past any already trained steps if resuming training
if steps_trained_in_current_epoch > 0:
steps_trained_in_current_epoch -= 1
continue
tr_loss += self._training_step(self.model, inputs)
if (step + 1) % self.args.gradient_accumulation_steps == 0 or (
len(epoch_iterator) <=
self.args.gradient_accumulation_steps and
(step + 1) == len(epoch_iterator)):
global_step += 1
if self.args.local_rank in [-1, 0]:
if (self.args.logging_steps > 0
and global_step % self.args.logging_steps
== 0) or (global_step == 1
and self.args.logging_first_step):
logs = {}
if self.args.evaluate_during_training:
results = self.evaluate()
for key, value in results.items():
eval_key = "eval_{}".format(key)
logs[eval_key] = value
loss_scalar = (tr_loss - logging_loss
) / self.args.logging_steps
learning_rate_scalar = get_lr_this_step(
global_step)
logs["learning_rate"] = learning_rate_scalar
logs["loss"] = loss_scalar
logging_loss = tr_loss
epoch_iterator.write(
json.dumps({
**logs,
**{
"step": global_step
}
}))
if self.args.max_steps > 0 and global_step > self.args.max_steps:
epoch_iterator.close()
break
if self.args.max_steps > 0 and global_step > self.args.max_steps:
train_iterator.close()
break
logger.info("\n\nTraining completed. \n\n")
return TrainOutput(global_step, tr_loss / global_step)
def _training_step(self, model: ORTTrainer,
inputs: Dict[str, torch.Tensor]) -> float:
for k, v in inputs.items():
inputs[k] = v.to(self.args.device)
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in transformers (see doc)
return loss.item()
def save_model(self, output_dir: Optional[str] = None):
output_dir = output_dir if output_dir is not None else self.args.output_dir
os.makedirs(output_dir, exist_ok=True)
self.model.save_as_onnx(os.path.join(output_dir, "transformer.onnx"))
def evaluate(self) -> Dict[str, float]:
"""
Run evaluation and return metrics.
Returns:
A dict containing:
- the eval loss
- the potential metrics computed from the predictions
"""
eval_dataloader = self.get_eval_dataloader()
output = self._prediction_loop(eval_dataloader,
description="Evaluation")
return output.metrics
def predict(self, test_dataset: Dataset) -> PredictionOutput:
"""
Run prediction and return predictions and potential metrics.
Depending on the dataset and your use case, your test dataset may contain labels.
In that case, this method will also return metrics, like in evaluate().
"""
test_dataloader = self.get_test_dataloader(test_dataset)
return self._prediction_loop(test_dataloader, description="Prediction")
def _prediction_loop(self, dataloader: DataLoader,
description: str) -> PredictionOutput:
"""
Prediction/evaluation loop, shared by `evaluate()` and `predict()`.
Works both with or without labels.
"""
logger.info("***** Running %s *****", description)
logger.info(" Num examples = %d", len(dataloader.dataset))
logger.info(" Batch size = %d", dataloader.batch_size)
eval_losses: List[float] = []
preds: np.ndarray = None
label_ids: np.ndarray = None
self.model.eval()
for inputs in tqdm(dataloader, desc=description):
has_labels = any(
inputs.get(k) is not None
for k in ["labels", "masked_lm_labels"])
for k, v in inputs.items():
inputs[k] = v.to(self.args.device)
with torch.no_grad():
outputs = self.model(**inputs)
if has_labels:
step_eval_loss, logits = outputs[:2]
eval_losses += [step_eval_loss.mean().item()]
else:
logits = outputs[0]
if preds is None:
preds = logits.detach().cpu().numpy()
else:
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
if inputs.get("labels") is not None:
if label_ids is None:
label_ids = inputs["labels"].detach().cpu().numpy()
else:
label_ids = np.append(
label_ids,
inputs["labels"].detach().cpu().numpy(),
axis=0)
if self.compute_metrics is not None and preds is not None and label_ids is not None:
metrics = self.compute_metrics(
EvalPrediction(predictions=preds, label_ids=label_ids))
else:
metrics = {}
if len(eval_losses) > 0:
metrics["loss"] = np.mean(eval_losses)
return PredictionOutput(predictions=preds,
label_ids=label_ids,
metrics=metrics)
def run_test(model, model_desc, device, args, gradient_accumulation_steps,
fp16, allreduce_post_accumulation, get_lr_this_step,
use_internal_get_lr_this_step, loss_scaler,
use_internal_loss_scaler, batch_args_option):
dataloader = create_ort_test_dataloader(model_desc.inputs_,
args.batch_size, args.seq_len,
device)
model = ORTTrainer(
model,
None,
model_desc,
"LambOptimizer",
map_optimizer_attributes=map_optimizer_attributes,
learning_rate_description=IODescription('Learning_Rate', [
1,
], torch.float32),
device=device,
postprocess_model=postprocess_model,
gradient_accumulation_steps=gradient_accumulation_steps,
# BertLAMB default initial settings: b1=0.9, b2=0.999, e=1e-6
world_rank=args.local_rank,
world_size=args.world_size,
use_mixed_precision=fp16,
allreduce_post_accumulation=allreduce_post_accumulation,
get_lr_this_step=get_lr_this_step
if use_internal_get_lr_this_step else None,
loss_scaler=loss_scaler if use_internal_loss_scaler else None,
_opset_version=12)
# trainig loop
eval_batch = None
model.train()
for step, batch in enumerate(dataloader):
if eval_batch is None:
eval_batch = batch
if not use_internal_get_lr_this_step:
lr = get_lr_this_step(step)
learning_rate = torch.tensor([lr])
if not use_internal_loss_scaler and fp16:
loss_scale = torch.tensor([loss_scaler.loss_scale_])
if batch_args_option == BatchArgsOption.List:
if not use_internal_get_lr_this_step:
batch = batch + [
learning_rate,
]
if not use_internal_loss_scaler and fp16:
batch = batch + [
loss_scale,
]
outputs = model(*batch)
elif batch_args_option == BatchArgsOption.Dict:
args, kwargs = split_batch(batch, model_desc.inputs_, 0)
if not use_internal_get_lr_this_step:
kwargs['Learning_Rate'] = learning_rate
if not use_internal_loss_scaler and fp16:
kwargs[model.loss_scale_input_name] = loss_scale
outputs = model(*args, **kwargs)
else:
args_count = int(len(model_desc.inputs_) /
2) # approx helf args, half kwargs
args, kwargs = split_batch(batch, model_desc.inputs_, args_count)
if not use_internal_get_lr_this_step:
kwargs['Learning_Rate'] = learning_rate
if not use_internal_loss_scaler and fp16:
kwargs[model.loss_scale_input_name] = loss_scale
outputs = model(*args, **kwargs)
print(outputs[0])
# eval
model.eval()
if batch_args_option == BatchArgsOption.List:
outputs = model(*batch)
elif batch_args_option == BatchArgsOption.Dict:
args, kwargs = split_batch(batch, model_desc.inputs_, 0)
outputs = model(*args, **kwargs)
else:
args_count = int(len(model_desc.inputs_) /
2) # approx helf args, half kwargs
args, kwargs = split_batch(batch, model_desc.inputs_, args_count)
outputs = model(*args, **kwargs)
return (output.cpu().numpy() for output in outputs)
